[CI] Fix debug_str() to be compatible with latest PyTorch nightly

[yf225]

PyTorch PR that changed stacktrace summary formatting: pytorch/pytorch#165397.

Example failing CI job: https://github.com/pytorch/helion/actions/runs/18923691664/job/54025942215?pr=1049

  def root_graph_0():
-    # File: .../basic_kernels.py:21 in torch_ops_pointwise, code: out[tile] = torch.sigmoid(torch.add(torch.sin(x[tile]), torch.cos(y[tile])))
+     # File: .../basic_kernels.py:21 in torch_ops_pointwise, code: out[tile] = torch.sigmoid(torch.add(torch.sin(x[tile]), torch.cos(y[tile])))
? +
      x: "i32[1024]" = helion_language__tracing_ops__host_tensor('x')
      block_size_0: "Sym(u0)" = helion_language__tracing_ops__get_symnode('block_size_0')
      load: "i32[u0]" = helion_language_memory_ops_load(x, [block_size_0], None, None);  x = None
      sin: "f32[u0]" = torch.ops.aten.sin.default(load);  load = None
      y: "i32[1024]" = helion_language__tracing_ops__host_tensor('y')
      load_1: "i32[u0]" = helion_language_memory_ops_load(y, [block_size_0], None, None);  y = None
      cos: "f32[u0]" = torch.ops.aten.cos.default(load_1);  load_1 = None
      add: "f32[u0]" = torch.ops.aten.add.Tensor(sin, cos);  sin = cos = None
      sigmoid: "f32[u0]" = torch.ops.aten.sigmoid.default(add);  add = None
      convert_element_type: "i32[u0]" = torch.ops.prims.convert_element_type.default(sigmoid, torch.int32);  sigmoid = None
      out: "i32[1024]" = helion_language__tracing_ops__host_tensor('out')
      store = helion_language_memory_ops_store(out, [block_size_0], convert_element_type, None);  out = block_size_0 = convert_element_type = store = None
      return None

[oulgen]

someone removed the space?

[yf225]

someone removed the space?

Yes I believe so - and this is probably a better fix: pytorch/pytorch#166590. Will close this PR instead.

[yf225]

@oulgen I realized Sherlock's change (pytorch/pytorch#165397) actually made the formatted graph code look nicer (no more misaligned indentation on the comment), so I will probably merge this PR instead